Confused about which tech stack to learn, what to prioritize, or how to actually break in? Book a real conversation with a working AI engineer — not another course, a person who answers your specific confusion.
Upload your resume and I'll personally go through it and send you honest, specific feedback you can act on.
You're at the centre. Hover (or tap) a tool to see what it's for and when it matters.
Learn from a working Generative AI engineer who has shipped this stuff. Book a 20-minute session on exactly what you're stuck on.
Helps beginners get unstuck on where to even start — picking the right tech stack, building a realistic roadmap, and avoiding the mistakes that waste months.
Not generic mentorship talk — pick the kind of session that matches exactly what you're stuck on.
Confused between Python vs other languages, TensorFlow vs PyTorch, LangChain vs LangGraph? I help you choose based on where you actually want to go, not by trying everything at once.
Bring your specific confusion — classical ML, RAG, LangGraph, MCP, whatever's blocking you — and leave with a practical, non-academic explanation.
Get a structured path built for your current level, and learn the mistakes to avoid — the ones that cost me months when I was starting out.
1:1 review of what you're building or applying with — direct, specific feedback instead of generic advice.
Built and deployed ML models on AWS SageMaker for forecasting and anomaly detection; automated ETL pipelines with Glue, Athena, and Lambda.
Building LangGraph-based agent flows, RAG-based knowledge retrieval, and MCP tool orchestration for enterprise AI assistant platforms.
I'm Rohan, from Asansol, West Bengal — now working in Generative AI, building agentic systems, RAG pipelines, and LLM orchestration. I started exactly where most of you are now: confused about Java, DSA, and what actually mattered.
A best friend nudged me toward ML instead, and that hands-on curiosity turned into a full career. Someone once cut through my confusion in ten minutes — I mentor because I want to do the same for the next person stuck where I was.
When I was buried in Java and DSA, close to giving up, a friend nudged me somewhere completely different.
"Forget the DSA grind for a bit. Start learning ML instead, and try implementing it on some real data. You'll definitely like it."— my friend, the reason I found AI
Something clicked that Java and DSA never quite did. He's the entire reason I found my way into AI — and the reason I mentor others today.
Before you even book a session, start here. Tap a topic to expand — each one covers what to focus on, what to skip, and real mistakes made along the way.
Understand why a model works before touching sklearn. Master Linear/Logistic Regression, Decision Trees, Random Forest, KNN, SVM, Naive Bayes, and K-Means — plus data cleaning, feature engineering, and evaluation metrics. Real-world AI work is 70% data handling, not modeling.
Don't obsess over hyperparameter tuning frameworks or deep mathematical derivations before you understand the base algorithm and have built a few projects.
I built models on messy datasets without properly understanding the data first — accuracy looked fine, but the model was learning noise. Spend more time understanding your data than tuning your model.
Perceptrons, activation functions, forward/backward propagation, and gradient descent — conceptually, before frameworks. Then CNNs, RNN/LSTM basics, and Transformers. Pick one framework (PyTorch is a good default) and go deep.
Don't try to master every architecture — GANs, Diffusion Models — before you're comfortable training a basic feedforward or CNN model end-to-end.
I tried learning too many architectures at once without finishing one complete project. Build one full pipeline — data to deployment — before chasing the next architecture.
Prompts, Chains, Memory, and Retrievers cover 80% of real use. RAG is the single most practical skill right now — learn chunking, embeddings, vector storage (FAISS/Chroma), and retrieval. Then tool/function calling.
Don't get lost in every integration LangChain offers. Learn the pattern once with one vector DB and one LLM provider — it transfers everywhere else.
I used pre-built chains without understanding what was happening inside them. Build a RAG pipeline manually at least once so you know what the abstraction is doing for you.
The shift from linear pipelines to stateful, cyclical workflows. Understand nodes, edges, state management, conditional routing, and human-in-the-loop checkpoints — critical for production agent systems.
Don't build complex multi-agent systems on day one. Start with a single agent, 2-3 tools, and a simple decision loop.
I underestimated state management — my agents lost context between steps. Get comfortable with how state flows through your graph before adding complexity.
Ongoing ideas and honest guidance from me — proof of what you can expect in a real session, not just a taste. Tap any entry to read it in full.
For the longest time, my biggest confusion wasn't a specific DSA topic — it was not knowing how much DSA I even needed as someone heading toward AI engineering. I watched tons of videos, solved hundreds of scattered problems, and still felt lost, because I had no structure.
This sheet is what finally fixed that. It's a week-by-week DSA roadmap — problems, difficulty levels, and curated video lectures, all sequenced so you're never guessing what to do next. It was originally shared with me by one of my seniors, a teacher who genuinely cared about students figuring this out properly.
Every week I talk to someone with 12 half-finished courses and zero shipped projects. The problem was never a lack of resources — it's decision fatigue. Pick one roadmap, ignore every other "must-watch" video for 30 days, and just build. You'll learn more from one messy finished project than from five polished course completions.
Fine-tuning feels like the "advanced" move, so beginners chase it early. In practice, most production use cases are better served by retrieval — it's cheaper, easier to update when your data changes, and doesn't require retraining every time something new comes in. Save fine-tuning for when retrieval genuinely can't solve your problem, not as a default first step.
New agent frameworks show up constantly, and it's tempting to chase all of them. My filter: does it have real production adoption, active maintenance, and does it solve a problem my current stack genuinely can't? If the answer to any of those is no, I skip it — trend-chasing costs more time than it saves.
More posts coming soon — I'll keep adding ideas, notes, and research breakdowns here as I post them.
Upload your resume — your first review is free. I'll personally go through it and send you honest, specific feedback.
Whether it's a concept, a roadmap question, or a project you're stuck on — fill this in and I'll get back to you.
You can also reach me directly — I read every message.
nextaiengineers@gmail.com ↗Stuck on something, have an idea, or just want to talk AI engineering? Reach out over email, Instagram, or LinkedIn — whichever's easiest for you.